Terms / Memory & State / Selective Recall

Selective Recall

The system's ability to retrieve only what's relevant rather than everything it knows. The difference between a useful memory and a noisy one.

Selective Recall is about intelligent filtering at retrieval time. A system that stores 10,000 facts but dumps all of them into context every time is worse than a system that stores 10,000 facts and retrieves only the 15 that matter right now. Selective Recall requires understanding the current task well enough to determine which stored knowledge is relevant — and equally importantly, which stored knowledge should be left out. Over-retrieval is just as costly as under-retrieval: it creates Prompt Bloat from the memory layer.

Example

An agent is asked to add a new API endpoint. A system with poor Selective Recall retrieves everything tagged "API" — 4,000 tokens of context including authentication patterns, rate limiting rules, database migration procedures, the API versioning strategy, and the deprecation policy. A system with good Selective Recall retrieves only: the existing endpoint pattern for this resource type, the current auth middleware setup, and the team's naming convention — 600 tokens. The focused retrieval costs less and produces better results because the model isn't filtering noise.

Selective Recall

More in Memory & State

Context Anxiety

Ghost Knowledge

Context Rot

State Loss

Context Continuity

Memory Fidelity

Provenance Clarity

Recall Precision